GEO: being cited inside the AI-generated answer, not just linked beside it.

How generative engines actually work

The GEO paper models a generative engine as a four-stage workflow: reformulate the user query, retrieve sources, summarize them, and generate a structured answer grounded in cited materials. Visibility depends on more than rank — it depends on whether your page is retrieved, selected as supporting evidence, and then cited in a way that occupies meaningful space in the final answer.

Crucially, the paper's table-level results show especially large relative visibility gains for lower-ranked pages in conventional SERPs. Generative systems can redistribute attention away from pure link authority and toward useful, well-supported passages. That does not mean classic SEO stops mattering — retrieval still precedes generation — but the marginal value of 'answer-worthy' passages can be higher than most teams expect.

What actually wins in GEO (per the research)

The GEO paper's strongest tactics are not classic keyword tricks. The methods that measurably increased source visibility:

• Adding citations to your own claims — primary sources, named research, traceable data.
• Relevant quotations — quotable passages that can be lifted verbatim into a generated answer.
• Statistics — numeric evidence, benchmarks, and original data.
• Fluency and clarity — well-written passages that compose cleanly into a larger answer.

The paper reports up to 37% subjective impression lift on Perplexity using the best methods tested. Notably, keyword stuffing — the tactic associated with low-quality historical SEO — performed poorly. GEO rewards evidence, quotability, and synthesis-friendliness.

What the engines themselves say

Google's position is conservative and important: SEO best practices remain relevant for AI features, there are no additional requirements for AI Overviews or AI Mode, and there's no need to create AI-specific schema or files. AI feature traffic reports inside the standard Search Console web data.

Microsoft's Bing AI Performance is the clearest operational GEO blueprint today. It exposes total citations, average cited pages, grounding queries, page-level citation activity, and visibility trends across Copilot, Bing AI summaries, and partner integrations. It explicitly recommends clarity, tables, FAQ sections, evidence-backed claims, freshness, IndexNow, and Bing Places accuracy.

OpenAI says ChatGPT search ranking depends on factors intended to help users find reliable information, and publishers should allow OAI-SearchBot and published IP ranges for inclusion. Referral traffic is trackable via utm_source=chatgpt.com. Perplexity's Perplexity-User agent fetches pages in response to user questions and is not used for training.

Measuring GEO across engines

Measurement is uneven but workable. A practical GEO KPI framework:

• Bing AI Performance — citation count, cited pages, grounding queries.
• Search Console web data — segment by query sets and pages likely surfacing in AI Overviews.
• GA4 referrals — filter for chatgpt.com, perplexity.ai, copilot.microsoft.com.
• Manual prompt testing — weekly tests of target queries across ChatGPT, Perplexity, Gemini, Claude.
• Branded search lift — a lagging but reliable signal of AI-driven awareness.

Traffic reality — smaller, but more valuable

First-party analytics studies found AI referrals up 13× from July 2024 to May 2025 in tech/software, with 27% lower bounce rates, 28% more time on site, and 20% more pages per visit. Semrush reports the average AI-search visitor converts at 4.4× the rate of a traditional organic visitor.

Independent Pew Research analysis tempers this: users encountering AI summaries clicked traditional results less often and clicked links inside summaries only rarely. The honest interpretation: AI replaces some traffic, redirects some, and upgrades some. Optimize for value per visit and visibility share, not raw volume.

Bot policy and legal positioning

Unlike classic SEO, GEO requires an explicit policy on AI bot access. Key controls:

• Google-Extended — governs some Gemini training and grounding outside Search, separate from Googlebot.
• nosnippet, data-nosnippet, max-snippet — preview-length and content-reuse controls.
• OAI-SearchBot — allow for ChatGPT search inclusion; the separate GPTBot governs training use.
• llms.txt and similar emerging standards — worth tracking but not yet broadly honored.

High-value publishers need an explicit position on each of these controls, weighted against licensing opportunities, litigation risk, and desired AI visibility. For most B2B tech vendors, the pragmatic default is: allow retrieval bots for inclusion, restrict training-only bots unless there's a deal in place.

Where GEO is heading: agentic commerce

Google's Universal Commerce Protocol (UCP) enables agentic actions in AI Mode and Gemini — including direct buying — while preserving merchant-of-record status and existing Merchant Center feeds. For ecommerce and tech vendors with direct-purchase models, GEO is evolving from citation optimization into transaction-path optimization. Worth evaluating now for teams with the technical stack to pilot it.